MildredN/deepbrainz-blind-spots

Blind Spots of DeepBrainz-R1-0.6B

This dataset captures blind spots of the DeepBrainz-R1-0.6B model: Model Link.

I generated 10 diverse prompts to test the model’s responses against expected answers. The prompts cover factual knowledge, math, language, reasoning, and commonsense questions. This helps highlight where the model may give incorrect or misleading outputs.

---

Prompts Tested

1. "What is the capital of Burkina Faso?"
2. "If 10y + 5 = 30, find y."
3. "Define the word 'bank'."
4. "Explain why photosynthesis is important."
5. "What does the Igbo word 'nwachukwu' mean in English?"
6. "What are the major ethnic groups in Nigeria?"
7. "Evaluate 3 + 5 - 7 * 10"
8. "Who wrote the book 'Americana'?"
9. "If you have two apples and I take away three, how many do you have left?"
10. "Why is the sky blue?"

---

How the model was loaded

Example code snippet used in Colab:

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "DeepBrainz/DeepBrainz-R1-0.6B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, device_map="auto", torch_dtype=torch.bfloat16
)

prompt = "What is the capital of Burkina Faso?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50)
generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(generated_text)

Observed Blind Spots

Category	Example Issue
Factual errors	Misnaming capitals or historical facts (e.g., Burkina Faso → "Nouakam")
Arithmetic & logic errors	Mistakes in solving simple math or equations (e.g., solving 10y + 5 = 30)
Context drift & verbosity	Repeating phrases and unnecessary long responses
Ambiguity handling	Difficulty with words that have multiple meanings (e.g., "bank")
Commonsense failures	Errors in everyday reasoning (e.g., taking more apples than available)

Suggested Fine-Tuning Dataset

To fix the errors observed, the model could be fine-tuned on a diverse dataset covering:

• Factual QA – Questions about countries, capitals, history, and general knowledge to improve factual accuracy.
• Math & Reasoning – Simple and multi-step arithmetic, algebra, and logic problems to improve calculation and reasoning skills.
• Commonsense Reasoning – Everyday scenarios and puzzles to handle practical logic errors (e.g., taking more apples than available).
• Word Definitions & Multilingual Terms – Definitions, synonyms, and words from less common languages to handle ambiguity and language gaps.

How to assemble/find such a dataset:

• Use existing public datasets.
• Crowdsourcing small additional datasets for gaps or uncommon languages.
• Scraping structured factual information from reliable sources like Wikipedia or educational websites.

Recommended dataset size: Around 50,000 to 100,000 high-quality examples, ensuring coverage across all categories (facts, math, reasoning, language).
This will give the model enough examples to improve its accuracy and reduce blind spots.